An Adaptive PCA-Like Asynchronously Deep Reservoir Computing for Modeling Data-Driven Soft Sensors.

Reservoir computing (RC) is a promising tool to build data-driven models, which has exhibited excellent performance in dynamical modeling area. Asynchronously deep reservoir computing (ADRC) is an improved version of RC. It generates more diverse dynamics in the reservoir than traditional RCs because of multi-layered structure and asynchronous information process. Reservoir size is a key factor to affect the performance of ADRC. However, the reservoir size of ADRC is very difficult to be determined. To solve this problem, an adaptive PCA (principal component analysis)-like method is proposed. This method promotes the useful reservoir neuron signals while suppresses the useless reservoir neuron signals. It is very similar to the function of PCA in data process. The training way of the adaptive PCA-like ADRC (APCA-ADRC) is derived based on ridge regression technique. The validity of the APCA-ADRC is tested by modeling the SO2 concentration in sulfur recovery unit. Experimental results show it is prominent in building soft sensors with high performance.